The book that made Zhi-Hua Zhou (周志华) a household name among Chinese AI students has a watermelon on the cover. Good melons and bad melons get sorted by color, root and the sound they make when tapped, and the watermelon (西瓜) stays the running example for everything from decision trees to support vector machines.
Students call it the watermelon book (西瓜书). That is the Zhou most people know.
The other Zhou has spent most of his career arguing with a sentence that almost nobody else bothers to question — the assumption that when a technique works, the technique is the thing.
Key takeaways
- Zhou is a professor at Nanjing University, an academician (院士) of the Chinese Academy of Sciences and a vice president (副校长) of the university. He founded the LAMDA group and was first dean of its School of Artificial Intelligence, established in 2018.
- In 2017 his group published gcForest, or deep forest (深度森林): a deep model built from decision tree ensembles rather than neural networks, on the claim that what makes deep learning work is layer-by-layer processing, in-model feature transformation and sufficient complexity — not backpropagation itself.
- Since 2016 he has argued for learnware (学件): a trained model plus a machine-readable specification (规约), so models can be matched and reused instead of rebuilt.
- His team released Beimingwu (北冥坞), described in a KDD 2024 paper as the first open-source learnware dock system. It takes submitted models without anyone uploading raw data.
- The thread connecting all of it is not a model. It is a refusal to let one successful technique harden into the only technique.
The argument hidden inside deep forest (深度森林)
By 2017 the phrase "deep learning" had quietly become a synonym for "deep neural networks". Zhou and his then-student Ji Feng (冯骥) published a paper at IJCAI 2017 — Deep Forest: Towards An Alternative to Deep Neural Networks — that took the synonym apart.
Their move was to separate the mechanism from the implementation. Three properties, they argued, carry most of the load in deep neural networks: layer-by-layer processing, features transformed inside the model, and enough model complexity. None of those requires differentiable modules.
So they built a deep model that satisfies all three out of decision trees. gcForest stacks cascade layers of random forests; each layer's class-probability vector is concatenated back onto the original features and fed to the next, and the cascade grows only while held-out accuracy improves. A multi-grained scanning stage slides windows over raw features, roughly playing the role of convolution or recurrence. The payoff is far fewer hyperparameters than a deep network, with depth set by the data rather than by an architect.
This matters more than the benchmark table. gcForest is a worked demonstration that a dominant implementation can be mistaken for a principle.
Learnware (学件): what happens after training ends
The question Zhou returned to is unglamorous: the world already holds an enormous number of trained models, so why does every new task start from zero?
His answer, laid out in a 2016 short paper in Frontiers of Computer Science, is learnware (学件): a well-trained model plus a specification (规约) describing what it is good at. The specification has a semantic part, in text, and a statistical part that sketches the model's statistical footprint — in his team's work, a reduced kernel mean embedding (RKME). The point of the specification is that it lets a system judge whether a model is useful without the developer handing over their original data.
Put enough of these in one place and the workflow inverts: instead of training, you search, retrieve, and reuse — singly or in combination.
Beimingwu (北冥坞): the dock, not the model
Ideas like this usually die as position papers. This one did not. Zhou's group built a system and published it at KDD 2024 in Barcelona: Beimingwu: A Learnware Dock System, with Zhi-Hao Tan and Hai-Tian Liu as equal-contribution authors. Version 1.0 dates to January 2024, and the code is open source.
The engineering is where the idea gets tested: Beimingwu generates specifications locally, screens submitted models for quality, searches across tabular, image and text learnwares, and never requires raw data to leave the owner's machine.
The paper also states the ambition plainly: once a dock accumulates millions of learnwares, tasks nobody ever trained a model for might be solvable by assembling existing ones. That "might" is doing real work in the sentence.
The textbook and the institution
None of this would carry weight if Zhou were a marginal figure. His Chinese textbook 《机器学习》 (Tsinghua University Press, 2016) is used across universities, and his Ensemble Methods: Foundations and Algorithms (CRC Press) reached a second edition in 2025.
Nanjing University's page credits him with two National Natural Science Awards (second class) and the IEEE Computer Society Edward J. McCluskey Technical Achievement Award (2019). He is a fellow of ACM, AAAI, AAAS and IEEE, and a foreign member of Academia Europaea. LAMDA, founded in 2007, became one of China's best-known machine-learning groups, and the university's School of Artificial Intelligence, with Zhou as its first dean, was among the first at a top Chinese university.
Why this runs against the grain
Zhou is not a large-model (大模型) skeptic in the popular sense. In a Nanjing University feature, he speaks of a degree of "superstition (迷信)" about large models: forcing one onto a task a small model already handles wastes compute and can backfire. His objection is to monoculture, not to scale.
That distinction is easy to miss. Most public argument about Chinese AI is about who has the biggest model. Zhou's question is what happens to a field when nearly everyone works on the same object.
According to Li An, chief scientist at BrainNet (脑机网), the economics will eventually force reuse onto the agenda anyway, because retraining near-identical models is the single cheapest form of waste to eliminate once compute stops getting cheaper.
Honest limitations
This profile leans on Zhou's own papers, his university's pages and his CV — the right source for what he built, a poor source for how widely any of it is used. Textbook readership and Beimingwu adoption figures I could not verify independently.
I also cannot tell you whether the learnware (学件) bet pays off. Deep forest (深度森林) has not displaced neural networks, and I found no evidence that it was ever intended to. The honest framing is that Zhou's contributions are a research programme plus a working system, not a victory.
Two things are deliberately left out: AI education policy, and any claim about how his administrative roles interact with funding. Both need sources I could not verify. Where titles conflict, I used the university's own page.
What readers can do now
- Read the IJCAI 2017 gcForest paper before reading commentary about it. The three-property argument is two paragraphs long and is the part worth arguing with.
- If you train niche models on private data, look at the Beimingwu repository and ask the practical question: could you write a specification (规约) for your model that lets someone else use it without seeing your data? If you cannot, you have found the hard part.